Predicting Long-Term Unemployment Risk; Andreas Mueller (The University of Texas at Austin)

Abstract

This paper uses rich administrative and survey data from Sweden to study the predictability and determinants of long-term unemployment (LTU) over the period from 1992 to 2016. We use standard machine learning techniques to predict job seekers' LTU risk and find substantial predictable heterogeneity. Compared to a prediction model, which only uses standard socio-demographic variables, a model that uses data on income, employment and benefit histories more than doubles the predictive power of LTU risk. Moreover, the estimated heterogeneity in LTU risk in the full model implies that at least two thirds of the observed duration dependence in job finding is driven by dynamic selection. Finally, we apply our prediction algorithm over the business cycle and find that, in recessions, LTU risk is more compressed at the bottom of the distribution, suggesting a lower value of targeting unemployment policies